1. No category

Sharing the Loves: Understanding the How and Why of Online Content Curation
Changtao Zhong
Sunil Shah
King’s College London
changtao.zhong@kcl.ac.uk
Last.fm
sunil@last.fm
Karthik Sundaravadivelan
Nishanth Sastry
King’s College London
karthik.sundaravadivelan@kcl.ac.uk
King’s College London
nishanth.sastry@kcl.ac.uk
Abstract
This paper looks at how and why users categorise and
curate content into collections online, using datasets
containing nearly all the relevant activities from Pinterest.com during January 2013, and Last.fm in December 2012. In addition, a user survey of over 25 Pinterest and 250 Last.fm users is used to obtain insights into
the motivations for content curation and corroborate results. The data reveal that curation tends to focus on
items that may not rank highly in popularity and search
rankings. Yet, curated items exhibit their own skewed
popularity, with the top few items receiving most of the
attention; indicative of a synchronised community. We
distinguish structured curation by active categorisation
from a more passive bookmarking by ‘liking’ an item,
and find the former more prevalent for popularly curated items. Likes, however, are initially accumulated at
a faster pace. Finally, we study the social value of content curation and show that curators attract more followers with consistent activity, and diversity of interests.
Interestingly, our user study indicates a divided opinion
on the relevance of the social network.
Introduction
Social content curation is a new trend which has emerged
following on the heels of the information glut created by the
user-generated content revolution. Rather than create new
content, content curation websites allow their users to categorise and organise collections of content created by others
that they find online. These users or content curators provide
an editorial perspective by highlighting interesting content.
Typically, a social component is also involved: users can follow other content curators that they find interesting, as a way
of gaining exposure to new and interesting content.
Although content curation has only recently become a
buzzword, sites on the Web have supported the actual process of categorising and sharing content with followers for
a few years now. For instance, delicio.us allowed users to
categorise interesting URLs by tagging them, and sharing
them with followers. digg.com and reddit.com have allowed
sharing of news articles, and so on. In this paper, we take a
broad view of content curation and seek to understand the
c 2013, Association for the Advancement of Artificial
Copyright Intelligence (www.aaai.org). All rights reserved.
basic process by examining two very different datapoints:
Pinterest, arguably the most popular content curation website for sharing pictures and videos, and Last.fm, a popular
social music recommendation service.
The two websites are used differently but have similar
support for content curation: Users on Pinterest.com ‘pin’
images onto boards to categorise them, or ‘like’ the images to express an interest without categorisation. Curation
is not the primary purpose of Last.fm, but relevant support
has existed for over seven years. Similar to Pinterest, users
can categorise music tracks by attaching tags to them, or
‘love’ tracks to express interest without categorising. Tags
on Last.fm can be used to create playlists or personalised radio stations. Both sites support social networks where users
can form social links, based on their interests.
Through these websites we seek to understand why people curate, how they curate, and what others, namely followers of the content curators, find useful. For our analysis, we
collected nearly all of the curation actions and the entire social networks of active users from both websites for a period
of time: for three weeks in January 2013 for Pinterest, and
over the month of December 2012 for Last.fm. To complement the patterns we observe from the collected data and
to better understand the motivations for content curation, we
conducted a user study through online survey questionnaires
and interviews, recruiting over 25 and 250 users of Pinterest
and Last.fm respectively.
To understand why people curate, we look at the popularity distributions of highly curated items. The most popular
curated items appear to be of niche interest that may not rank
highly in other popularity rankings. For instance, the items
most pinned or most liked on Pinterest are from websites
with a low PageRank value or Alexa Global Traffic Ranking.
Similarly, the most tagged or loved tracks on Last.fm tend to
have a low rank in weekly music charts of total radio airplay, and greatest sales volume. We conjecture that curation
might provide a personal value to the curators by collecting
together items which may be difficult to find by other means.
Our user studies provide support to the notion that curation
provides personal value to the curators. Interestingly, despite
their low popularity in other rankings, there appears to be a
consensus on which items are most curated, and curation
actions are highly skewed towards the top items on each
site: The top 10% (0.1%) of items get over 70% of the cura-
tion actions on Pinterest (Last.fm), indicative of a synchronised community. These findings provide evidence for Clay
Shirky’s theories that “curation comes up when search stops
working”, and that “the job of curation is to synchronize a
community so that when they’re all talking about the same
thing at the same time, they can have a richer conversation
than if everybody reads everything they like in a completely
unsynchronized or uncoordinated way” (Shirky 2010).
Next, we examine the different curation actions to better understand the process. Based on the similarity between
the curation actions on Pinterest and Last.fm, we propose a
distinction between two kinds of content curation actions:
unstructured curation, which involves highlighting or collecting interesting content without categorising them (e.g.,
‘like’ or ‘love’), and structured curation, categorising content along with other “similar” items from some perspective
(e.g., tag or ‘pin’). We find that different users prefer different actions, with some preferring unstructured, and others
structured curation. The proportion of users preferring each
varies from site to site, and may be a result of differences in
the way the site is structured and side effects associated with
each action (e.g., the action of ‘love’ing a track on Last.fm
is used to recommend other similar tracks to the user). However, ranking items based on the number of unstructured or
structured curation actions, we see that the top items in both
rankings receive more structured curation actions than unstructured. In contrast, for all items, we see that the easier
action of unstructured curation accumulates faster.
Finally, we study the social value of content curation.
Bhargava, who appears to have coined the term content curator, defined it as “someone who continually finds, groups,
organizes and shares the best and most relevant content on
a specific issue online. The most important component of
this job is the word ‘continually.’ ”(Bhargava 2009). Consistent with his view, we find that curators who are regular and
consistent in their activities accumulate the most number of
followers on the respective websites. Diversity of interests is
also similarly rewarded: Curators with an expertise in multiple genres of music in Last.fm or categories on Pinterest
are similarly successful in attracting followers. Additionally,
we find that in Pinterest, successful users are those who prefer structured curation or pinning to merely ‘liking’ items.
On Last.fm, where the relative value proposition to the user
of tagging and loving tracks are structured somewhat differently, we do not find the same advantage for structured
curation.
Methodology and Data Description
We used two complementary approaches to examine content curation on both the Pinterest and Last.fm websites.
The main approach was a quantitative analysis of datasets
containing relevant curation actions over a fixed period of
time. Our data-based findings were corroborated and complemented using a qualitative approach consisting of user
studies and interviews.
Data and website description
We first provide a background about each of the websites
we use, and the data we collect about their actions and their
social relations.
Pinterest Pinterest is a photo sharing website that allows
users to save images and categorize them on different collections. Images added on Pinterest are termed pins; we will
use the terms pin and image interchangeably. A pin can be
created by pinning or importing from a URL external to pinterest.com, or repinning from a existing pin on pinterest.
Users organise their pins into collections called pinboards or
boards. A board needs to be specified at the time of pinning;
pins may be moved to a different board later on. A repin creates a new pin on the repinning user’s board. Each board can
belong to one of 32 globally specified categories on pinterest. Each category has a page on pinterest.com, highlighting
the latest pins. In addition to pinning or repinning, users can
like a pin or comment on a pin. Likes express an interest in or
appreciation of a pin without adding it onto the liking user’s
collections. The most recent likers of a pin are listed on the
pin’s webpage on pinterest, and the likes of a user are collected on the user’s profile. In addition to these content curation actions, users can also actively follow other users or
boards they find interesting, effectively creating a directed
social graph.
To analyze the curation activity on Pinterest, we collected
nearly all activities by crawling the main site between 3 and
21 Jan, 2013. The crawl proceeded in two steps: firstly, to
discover new pins, we visited each of the 32 category pages
every 5 minutes, and collected the latest pins of that category. Secondly, for every pin obtained this way, we visited
the webpage of the pin every 10 minutes. A pin’s webpage
lists the 10 latest repins and the 24 latest likes; we added
these to our dataset, along with the approximate time of repins, likes and comments (if any). In this paper, we focus on
repins and likes which comprise the vast majority of actions.
For any pin, if more than 10 repins or 24 likes had accumulated since our last visit, we may have missed some data.
The danger of missing data is higher for popular images
which may accumulate likes and repins faster than other images. However, if we find an overlap between the latest repins/likes on successive visits, then we can be sure of not
having missed data. In practice, we find that even for popular
images (those with more than 500 actions), we have missed
data in less than 0.06% of visits for repins and 0.02% for
likes. For all images, the fraction of visits which resulted in
missed data stands at 5.7 × 10−6 for repins and 9.4 × 10−7
for likes.
In addition to these curation actions, we also obtained the
social graph of Pinterest using a snowball sampling technique starting from a seed set of 1.6 million active users
which we collected initially. In total 30.5 million users and
315.2 million directed edges between them were obtained.
Users with a local clustering coefficient of 0 were filtered1
1
This affects Pinterest results reported per-user (mainly in the
final section, “What Other People Find Useful”). We justify this
filtering on the basis that users with clustering coefficient of 0 have
far fewer activities on the site. For example, in our seed set of 1.6
million users, the average number of pins for users with zero clustering coefficient is 324.3, while this value for users with non-zero
clustering coefficient is 1686.3.
resulting in a smaller social graph of around 7.1 million
users and 192.7 edges. For each of the remaining 7.1 million users, we collected statistics such as a string description,
their boards, total number of pins and likes since joining the
site, and numbers of followers and following users.
Summary Table 1 provides a summary of the aggregate
volume of data collected, and Fig. 1 provides an indication
of the per-user distribution of the volume of the data.
User study
A qualitative approach was undertaken through user surveys
and semi-structured interviews. We questioned users of Pinterest and Last.fm on their general behaviour on each website and on their attitudes towards usage of various social
signals on each website. These signals were primarily liking
Pinterest
Last.fm
03–21 Jan 2013
8,452,977
96,390,143
19,907,874
38,041,368
01–31 Dec 2012
291,562
5,887,159
89,338,529
59,622,487
Table 1: Dataset details
1.0
1.0
0.8
0.8
0.6
Pins-Pinterest
Likes-Pinterest
Tags-Last.fm
Loves-Last.fm
0.4
0.2
0.0 0
10
101
102
103
104
Number of Actions per User
(a) Actions per user
105
CDF
CDF
Last.fm Last.fm is a popular social music recommendations website which offers a radio service, a service to allow users to submit and track what music they have listened
to and music recommendations calculated using collaborative filtering algorithms. The Last.fm social graph is a simple
symmetric directed graph where users can friend each other.
The friendship must be approved by each user. While there is
no explicit following model through which content curated
(i.e. loved or tagged) by friends is made visible to users,
users can see friends’ loved tracks via a link on their (personalised) Last.fm home page. A user’s tagging and loving
activity can also be seen by visiting their profile page. The
action of submitting a track’s name or identifier to Last.fm to
record a listen is known as a scrobble. All scrobbled tracks
are shown on the user’s profile and are public unless deleted.
Charts of their most scrobbled tracks are shown on their profile page. Users can also love a song - either retroactively via
the Last.fm website or via any of the Last.fm client applications. Loved tracks are shown on their profile page. Similarly, users can ban a song when listening to it via Last.fm
radio, either via the website or via any of the client applications. Bans are one of the most infrequently used social actions and although they are public, are not displayed on the
profile page. All three actions are used to influence playlisting for Last.fm radio as well as music recommendations.
Additionally, users can tag a song via the Last.fm website.
Tags are both global and local - any tags a user has applied
will be shown on their profile page, as well as aggregated
with other users’ tags and shown on artist and track pages.
Users can also comment on nearly every content page (catalogue and user) on the Last.fm website via the page’s shoutbox. However, due to the complexity involved in extracting
shouts, we have excluded them from our dataset.
Our dataset was generated directly from Last.fm source
data using Apache Hive running on a Hadoop cluster. We
consider all users of Last.fm worldwide that have both
tagged and loved tracks during December 2012. We filter
down this dataset to include all users who are have at least
one other friend within the same dataset. This yields nearly
300,000 users who we consider to be power users of the
site. For each of these users, we extract all stored scrobble,
love and tag data along with approximately 5.9 million undirected edges between users.
Timespan
Users
Relationships
Likes/Loves
Repins/Tags
0.6
0.4
Followers - Pinterest
Following - Pinterest
Friends - Last.fm
0.2
0.0 0
10
101
102
103
104
Degree
105
106
107
(b) Degree distribution
Figure 1: Distribution of the number of curation actions and
social relationships per user in Pinterest and Last.fm datasets
and pinning/repinning for Pinterest, and loving and tagging
for Last.fm.
Other key areas of interest included the motivation for
users to curate content and the underlying relationship between curation and social networking. Users were also questioned on the value they placed on the social aspect of each
site and if this went beyond the scope of purely curation activities.
Pinterest The users who participated in the Pinterest study
were sourced from the social circles of the authors and from
posting requests on a Pinterest forum known as Pintester
and on the official Pinterest Facebook page. 30 users participated in the survey and a further 3 were interviewed. Interviews were loosely structured around the survey aims and
were presented to interviewees in themes. Interviewees were
asked to narrate their behaviour given certain scenarios and
were allowed to continue on their motivations and reasons
behind said behaviour. They were also encouraged to provide examples of their past practices and suggest how and
why these behavioural trends varied, if at all. Our main interest was to understand user perceptions and to find motivations behind their curating activities. Insight into how they
perceived social connections within the site was also sought
after. Interview responses were coded and then compared
against trends prevalent in the survey responses.
Last.fm A much larger user population of 270 participated
in the Last.fm survey since the survey was advertised over
the official Last.fm Twitter account2 . The questions posed
to users followed the structure adopted by the Pinterest survey but were specific to the dynamics and infrastructure
of Last.fm. Unstructured responses were obtained through
open-ended questions and were analysed to provide reasoning behind predominant user tendencies.
2
The study itself was carried out independently of Last.fm.
Why People Curate
In this section, we seek to find implicit reasons for why people curate by examining the characteristics of the content
they curate. Our approach will be to compare different popularity ranks with basic ranks created by the volume of curation actions. First, we ask where the content in curation system is from, by correlating curation with traditional popularity ranks, and show that curation serves a different purpose
than, say, search. Then, the distribution of curation activity
is analysed and a highly skewed distribution is obtained, revealing that users synchronise and focus on the same small
number of items. We draw on our user studies to provide
support for and comment on these findings.
Curation highlights new kinds of content
A first question is whether curation serves a new and different purpose from other approaches to finding and highlighting interesting content. Popularity rankings traditionally
highlight content which a community finds useful. Therefore, we compare curation with other traditional notions
of popularity. In the case of Last.fm, we compare against
weekly sales and radio airplay charts published by Music
Week3 , a trade paper for the UK record industry and an established music data provider. In the case of Pinterest, we
do not have a well accepted global popularity ranking of images. As a proxy, we use the website where the curated image was originally found, and compare the rank of a website
on Pinterest (in terms of number of repins and likes), with its
rank in search (PageRank value, obtained from Google via
its Search API), and its global traffic ranking (according to
Alexa4 ).
Avg. Repins
Avg. Likes
Alexa Ranking
PageRank
Avg. Repins
Avg. Likes
/
0.912
-0.010
0.195
0.912
/
0.032
0.150
Table 2: Curation highlights websites not popular in
other rankings. Low correlation coefficients between
curation-based ranking of websites (ranking by the average
number of repins or likes) and traditional websites rankings (Alexa Traffic Ranking and Google PageRank) reveal
that curation serves a new purpose of highlighting nontraditional sites.
Loves
Tags
Music Week Sales
Music Week Airplay
Loves
Tags
/
0.39 (0.74)
-0.02 (0.05)
-0.11 (-0.01)
0.39 (0.74)
/
0.04 (0.14)
0.04 (-0.04)
Table 3: Curation highlights tracks not popular in traditional rankings. Low correlation coefficients between
curation-based ranking of tracks (ranking by number of tags
or likes) and traditional music track rankings (obtained from
Music Week) reveal that curation may be being used to find
music that is “off the charts” (i.e., is not mainstream). The
coefficients shown consider UK-based users only for the
curation-based rank, for a fair comparison with Music Week,
a UK-based rank. Numbers in brackets indicate correlation
coefficients with a ranking considering users worldwide.
Shirky’s theory that “curation comes up when search stops
working”(Shirky 2010). Similarly, Table 3 shows a similar
lack of correlation between highly ranked tracks through curation and the traditional Music Week rankings.
Curation for personal vs. social value
A second aspect of Shirky’s theory is that the “job of curation is to synchronize a community so that when they’re all
talking about the same thing at the same time, they can have
a richer conversation than if everybody reads everything they
like in a completely unsynchronized or uncoordinated way”.
We find evidence for this by examining the distribution of
curation actions in our corpus. Fig. 2 shows a highly skewed
popularity distribution, with a large proportion of the user
base curating a selected minority of items. However, that
skewness is expected in popularity distributions, hence this
is not in itself a confirmation of a community which consciously synchronises itself.
Repins
Top 1% Pins
5%
10% 20%
Likes
Top 0.1% Tracks
Tags
0.5%
Loves
In the case of Pinterest, we find that websites with highly
repinned or liked images tend not to have a high PageRank
or Alexa Global Traffic Rank. In fact, Table 2 shows that,
when considering all websites, there tends not to be a correlation between ranking based on number of repins/likes
and traditional ranking based on Google PageRank or Alexa
Global Traffic estimates. Thus, we conclude that curation
highlights a different set of sites compared to search and traffic. The low correlation with PageRank also lends support to
3
4
http://www.musicweek.com/
http://www.alexa.com
0.0
0.2
0.4
0.6
0.8
1.0
Figure 2: Distribution of curation activity is highly
skewed towards a few popular items. In Pinterest, nearly
40% of curation activities (Repins and Likes) is for top
1% of images and over 73% are for top 10% images. In
Last.fm,the top 0.5% of tracks account for about 90% of curation activities (Tags and Loves).
Rather than answer the difficult question of whether curation creates value for users by synchronising a community,
How People Curate: Understanding Curation
Actions
As detailed before in the Methodology section, multiple curation actions are available to a user. For instance a user
on Pinterest can pin an item, like it or comment on it. On
Last.fm, tracks can be tagged, loved or banned or shouted.
We note that these actions can be distinguished based on
whether they simply highlight an item (love, like, ban, comment, shout), or they also organise the item onto userspecific lists (pinning an item onto a user’s board, or attaching a user’s tag to a track). We term the former as unstructured curation and the latter structured curation because of
the organisational structure induced by pinning or tagging.
We use this framework to study curation actions: Do users
have a preference towards one kind of action, do they use
structured actions preferentially in one setting, what the relative dynamics of the different kinds of action are, etc.
To investigate the relationship between the two forms of
curation, we define an unstructured curation ratio R as:
U nstructured
U nstructured + Structured
R=
(1)
Some users prefer structured, others unstructured
First we explore how users curate content, and whether they
prefer structured or unstructured curation. We calculate the
unstructured curation ratio R for each dataset and consider
the top 1%, the top 10% and all users for each activity on
both websites in Fig. 35 .
We define users who prefer structured curation over unstructured curation (i.e., have R < 0.5) as structured curators. Conversely, users who prefer unstructured over structured curation (R > 0.5) are termed unstructured curators.
1.0
1.0
0.8
0.8
0.6
0.6
0.4
Top 1% by Pins
Top 10% by Pins
Top 1% by Likes
Top 10% by Likes
All Users
0.2
0.0
0.0
0.2
0.4
R
0.6
0.8
(a) Users on Pinterest
1.0
CDF
CDF
we turn to our user study to find if users perceive their community to be useful (and thereby, determine whether the social value of curation is a motivating factor for why people
curate). Some users value the ability to talk with others who
share a similar taste. For instance, one Last.fm user states:
Social connection with people who share a similar taste
in music is an exciting thing. Say, for example, that
you’ve found a band that you love, and you live in an
area where people generally prefer listening to pop or
maybe classical or filmi (Bollywood) tracks. So, you
start looking for people around you who share the same
taste as you in music and become fast friends. Last.fm
has simplified that process.
A Pinterest respondent values the ability to serendipitously
discover through other users’ items which they might like,
placing an implicit value in the Pinterest community:
I like the feeling of stumbling on things which I did not
know I would like but I do.
However, such views are from a minority of users. A number of users use curation sites as a personal tool: 85% of
Pinterest respondents use it as a personal collection or scrapbook and only 48% of the population use the site to display
their content to others (Note that our survey allowed multiple answers to be selected for this question). 39% of Last.fm
users tag tracks for personal classification, whereas another
39% tag to create a global classification according to genres.
The majority of users shared in this Last.fm user’s view:
I find the social aspect more useful and interesting with
people I know, rather than developing new interactions
based on music taste.
One Pinterest user felt strongly about their aversion towards social interaction on the site:
I don’t really see a point (in communicating with a fellow user). And also the beauty of Pinterest, is the ability
to pin things from strangers. Why would I want to get
to know them.
Thus, we conclude that although the community of users
may focus its curation actions on a few items (as seen from
the popularity skew), this synchronisation is not a conscious
effort. Users, largely, are not actively trying to curate for social value and do not try to integrate within their respective
communities.
All Users
Top 1% by Loves
Top 10% by Loves
Top 1% by Tags
Top 10% by Tags
0.4
0.2
0.0
0.0
0.2
0.4
R
0.6
0.8
1.0
(b) Users on Last.fm
Figure 3: CDF of users’ unstructured curation ratio R In
both Pinterest and Last.fm, there are a mixture of structured
(R < 0.5) and unstructured (R > 0.5) curators. However,
generally Pinterest users participate in structured curation
activities whilst most Last.fm users participate in unstructured curation activities.
In Fig. 3 we first draw attention to the difference between the proportion of structured and unstructured curators on each network. Fig. 3a shows that on Pinterest, more
than 80% of all users are structured curators. Comparatively,
Fig. 3b shows that less than 40% of all Last.fm users are
structured curators. This corresponds with the findings of
our user study. We found that the majority of Pinterest users
surveyed would rather repin a post than like it if it matched
their interests. This is irrespective of whether the post was
from a user they were following or not. Similarly, a majority
of Last.fm users would rather love than tag a music track.
However, as expected, when filtering for the top 1% and
10% of users for each curation activity, we see that the unstructured curation ratio moves closer in favour of that activity, on both websites: The most prolific likers on Pinterest
are unstructured curators (i.e., R > 0.5 for these users, despite the prevalence of pinning on Pinterest); the most prolific taggers on Last.fm are structured curators (i.e., R < 0.5,
despite the importance of loves on Last.fm).
5
Since Pinterest does not distinguish between original pins and
repins, both are included to represent the structured curation action.
The larger proportion of the top users by loves who are
unstructured curators on Last.fm can be explained by the
relative prevalence of loves to tags in our dataset, as well
as one of the major side effects of loves. When a user loves
a track on Last.fm, this action is fed back into their music
recommendations and displayed to their friends. Loves are
thus a more capable curation activity on Last.fm compared
to likes on Pinterest. This is confirmed by our user study:
65% of surveyed Last.fm users have never tagged a track.
Conversely, only 11% have never loved a track.
1.0
CDF
0.8
0.6
Repins - 5th
Likes - 5th
Repins - 30th
Likes - 30th
Repins - 500th
Likes - 500th
0.4
0.2
0.0
0
50
Structured curation is preferred for popular items
1.0
CDF
0.8
0.6
0.4
Top 1% Repinned Pins
Top 1% Liked Pins
Top 1% Tagged Tracks
Top 1% Loved Tracks
0.2
0.0
0.0
0.2
0.4
R
0.6
0.8
1.0
Figure 4: CDF for unstructured curation ratio R of top
items on Pinterest and Last.fm Magenta line indicates
R = 0.5. In Pinterest and Last.fm, all of the top items have
R < 0.5, i.e. they are all subject to structured curation. Notice that even the top items for unstructured curation (i.e.,
top liked or loved items) have R < 0.5.
Unstructured curation is faster than structured
In this section, we discuss how items accumulate different
curation activities over time. In order to compare these, we
plot the action time - the time span between the n-th action
and the time a content item was originally posted. We consider this time for both structured (pin/tag) and unstructured
(like/love) curation activities.
Fig. 5 shows the time taken for items to reach their 5th,
30th and 500th curation actions on Pinterest. We find that
the majority of pins reach 5 curation actions (whether repin
150
200
250
0.40
0.35
0.30
0.25
0.20
0.15
0.10
0.05
0.00
KS
Figure 5: CDF of Pinterest repin and like action times The
kth repin (like) time for a pin is the time between creation of
a pin and the kth repinning (liking) in Pinterest. Likes accumulate quicker at first and there is a considerable difference
between in the time it takes to get 30 repins and 30 likes. The
distributions of the times for kth likes and repins converge
as k increases to 500.
KS
Next we explore how items themselves are curated, and
whether the majority of items are curated in a structured
or unstructured manner. We calculate the unstructured curation ratio R for each item in both datasets and consider the
top content items by curation activity in Fig. 4. We observe
that regardless of the ranking method used (i.e., whether the
ranking is based on the volume of structured or unstructured curation action received), the majority of items have
an R < 0.5: there are more structured curation actions for
top items, whether they are the top items for structured or
unstructured curation. In other words, even top liked items
have more pins than likes on Pinterest (similarly for Last.fm,
top loved items have more actions adding tags than actions
‘loving’ the track). This is further supported through our
Pinterest user study where average R for popular content
was 0.33 and for unpopular content was 0.5. The Last.fm
survey did not address this question.
100
Action Times(Hour)
0
100
200
300
kth time
(a)
400
500
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
50
100
150
kth time
200
250
(b)
Figure 6: Kolmogorov-Smirnov statistic for action times
Extending Fig. 5, for both Pinterest and Last.fm, there is
a difference between structured and unstructured curation
over successive actions. There is a noticeable peak for Pinterest, after which the difference between the two actions is
minimised. For Last.fm, we see a similar tail off, albeit after
many more actions.
or like) in several hours. As expected, it takes much longer
to reach their 30th curation action. However, there is a considerable difference between the 30th action time for likes
vs. repins: For 80% of items, accumulating 30 likes take approximately 100 hours whilst repins take approximately 200
hours. This difference decreases when we consider the 500th
action times for each activity.
In Fig. 6, we summarise the difference between the distribution of Ts (k), the kth action time for structured curation, and the distribution of Tu (k), the kth action time
for unstructured curation. This difference can be measured
using the Kolmogorov-Smirnov statistic given by D =
max (Ts (k) − Ts (k)). For Pinterest, we see a quickly growing difference between likes and repins until a initial peak,
after which the two converge again - suggesting that, initially, likes accumulate faster than repins. As items become
more popular, repins catch up and the two grow at a similar
rate. For Last.fm, we see a similar result - except that structured curation activity (tagging) does not completely catch
up with unstructured curation activity (loving). This can be
explained by what we show in Fig. 4 - structured curation is
generally stronger for Pinterest items than those on Last.fm.
Consistent and regular updates
2
5
Number Of Listens Before Action
>5
Figure 7: Number of listens before first love and tag on
Last.fm reveals that most users who curate an item do so on
their first listen.
A basic question that arises is when does a curation action happen during the content consumption cycle of a user.
While it is difficult to answer this question in general, our
data for Last.fm includes ‘scrobbles’ which record each listen. Using this, Fig. 7 shows the number of activities that
a user participates in before loving or tagging a track on
Last.fm for the most popular 1,000 tracks in our December
dataset. Interestingly, we find that most users who love or
tag a track predominantly on their 1st listen. A insignificant
minority of users listen to a track twice before loving or tagging it and no user takes longer than 2 listens to tag a track
and 5 listens to love a track.
What Do Other People Find Useful?
Although as suggested previously, many users view curation
as a highly personal activity, some users accumulate more
followers than others. This section sheds light on what curation behaviours other people find useful by using the number of accumulated followers as a metric. In each case, we
consider the per-user distribution of the values for some attribute of the user’s behaviour (e.g., interval between repins,
number of music genres the user is interested in, or the unstructured curation ratio R). Firstly, we separate users into
bins. Usually, we do this based on the user’s value of the
attribute considered (e.g., based on the board categories of
the user). Next for each bin, i.e., for each value of the attribute being considered, we compute the 90th percentile of
the number of followers accumulated by users in the bin as
a measure of how useful the bin’s value of the attribute is, to
other users.
In summary, for both Last.fm and Pinterest, we find that
regular curators who have a short interval between successive curation actions accumulate more followers, as do curators who have a diversity of interests. In Pinterest, we also
find that users who prefer structured curation (i.e., those who
prefer ‘pinning’ to ‘liking’) accumulate more followers. This
result does not carry over to Last.fm, where structured curation does not have the same predominant role. We have verified that each of the results in this section are robust against
the choice of 90th percentile as a summary measure. (Similar results hold for 80th and 50th percentile values as well).
300
280
260
240
220
200
180
160
Friends (90th Percentile)
1
1 hour 6 hours12 hours1 day 2 days 3 days 1 week 1w3d 2 weeks
60
50
40
30
20
10
0
1 hour
Repinning Intervals
(a) Pinterest
6 hours
12 hours
1 day
2 days
Tagging Intervals
3 days
1 week
(b) Last.fm
Figure 8: Structured curation attracts followers when it
is consistent and regular Users with a short interval between successive repins (tags) attract a large number of followers (friends) on Pinterest (Last.fm). A similar result can
be obtained for unstructured curation (loves/likes) as well
(not shown).
Diversity of interests
4000
3500
3000
2500
2000
1500
1000
500
0
0
5
10
15
20
Categories
(a) Pinterest
25
30
35
Friends (90th Percentile)
Loves
Tags
Bhargava has suggested that the most important part of a
content curator’s job is to continually identify new content
for their audience (Bhargava 2009). Fig. 8 examines the role
of regularity, by plotting the 90th percentile of the intervals
between consecutive structured curation actions6 for each
user vs. the 90th percentile of the followers accumulated,
and finds support for this theory. Note that for Pinterest,
too short an interval between repins could detract followers. However, Last.fm does not exhibit this phenomenon. We
conjecture that given the order of magnitude higher volume
of curation actions on Pinterest (See Fig. 1a), followers on
Pinterest may see too many repins as spam. Thus, Pinterest
users must not only be consistent and regular but must also
filter content by curating only the most interesting, in order
to attract followers.
Followers(90th Percentile)
16
14
12
10
8
6
4
2
0
Follower(90th Percentile)
Actions (x 100,000)
When do people curate?
250
200
150
100
50
0
0
20
40
60
Genres
80
100
(b) Last.fm
Figure 9: Diversity of interest attracts followers (a) Pinterest users interested in nearly all the categories attract more
followers; (b) Last.fm users interested in most of the music
genres tend to attract more followers.
Next, we examine the role of diversity. We capture diversity of a user’s interest in Pinterest by counting the number
of distinct categories (of the 32 globally recognised ones)
that the user has boards in. Similarly, for Last.fm, we capture diversity of interest by counting the number of genrespecifying tags which have been used for tagging by the
6
Because a user typically has many intervals between repins,
we additionally use a 90th percentile method when selecting this
attribute. That is, if a user’s structured curation intervals are represented as a list of intervals, I, this user will be put into a bin
according to the 90th percentile value of I.
user. Genre-specifying tags were identified by selecting the
top 150 tags on Last.fm and manually filtering out 22 nongenre tags, resulting in 128 distinct genre-specifying tags.
Fig. 9 shows that users who have an extremely diverse interest attract a large number of followers. However, beyond a
point, the number of followers falls off, for jack-of-all-trade
curators who are interested in nearly all categories or genres.
Note that there might be potential confounding factors:
For example, being active in a number of categories might
simply be a consequence of being more active on the site,
and more active users might attract more followers, as shown
above. To confirm that our finding about the importance of
diversity of interests is not simply an artifact of diversity in
usage, we verified that the result of Fig. 9 holds even when
we observe limited subsets of users with similar numbers of
pins (e.g., 1,000–2,000 pins, or 10,000–20,000 pins).
Structured vs. Unstructured Curation
300
280
260
240
220
200
180
160
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
R
Friends (90th Percentile)
Followers (90th Percentile)
In previous sections, we discussed structured and unstructured curation, and demonstrated that on Pinterest, most
users would prefer to use structured curation, whereas the
opposite is true for users of Last.fm. In this section, we try
to find out which kind of curation action is more useful for
other people.
In Pinterest, as shown in Figure 10, we find that with the
increase of unstructured curation ratio R, the numbers of followers decrease. This shows that structured curation (repin)
is more useful to others.
However, we do not observe a similar trend in Last.fm.
We hypothesise that this is because repinning is the dominant curation method in Pinterest, but tagging is not in
Last.fm. On the contrary, as explained previously, Last.fm
users are rewarded for ‘loving’ a track because Last.fm recommends other tracks which might be interesting to the
user. Thus, unstructured curation is much more prevalent in
Last.fm; even users who tag extensively also use ‘love’s, increasing their R ratios.
50
Qualitative studies
Some researchers have qualitatively studied the phenomenon of curation on online social networks. For example, (Liu 2010) identifies seven curatorial activities (collecting, organizing, preserving, filtering, crafting a story, displaying, and facilitating discussions) based on an analysis
of 100 web artefacts, and introduces the concept of sociallydistributed curation, to emphasize the distributed nature of
this curatorial process emerging from the social Web. (Rotman et al. 2012) explores the opportunities and challenges
of creating and sustaining large-scale content curation communities through a case study of the Encyclopedia of Life
(EOL)7 . A qualitative study among the personnel of a newspaper (Villi 2012) indicates that engaging the audience in
social curation is more important than involving the audience in content production.
However, without taking users’ behaviour and attitude
into consideration, why and how curation happens is still
unclear. Therefore, in our user study, the motivation and
process of curation are analysed. (Duh et al. 2012) also
looks into motivations for curation, by manually inspecting
435 lists of Tweets curated on Togetter.com and identifying
seven use cases for curation. Our paper takes a complementary approach, using orders of magnitude more data, and also
directly surveying users, obtaining new insights.
Quantitative studies
40
30
20
10
0
maintaining and adding value to, a trusted body of digital
information for current and future use” (Beagrie 2008).
In the Web, the idea behind curation is to link and/or excerpt the work of others (Carr 2012). Therefore, to some extent, the Web has always been about curation (Ovadia 2013),
with users sharing links with each other. Many blogs can be
considered as curated content, with bloggers sharing links
and excerpts with readers. Platforms like Pinterest, Tumblr
and Storify make it easier for users to share, showcase, and
curate content they discover online. As such, these sites give
an opportunity to deeply study the phenomenon of content
curation.
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
(a) Pinterest
R
(b) Last.fm
Figure 10: Structured curation attracts followers Users
with a low unstructured curation ratio R (i.e., those with a
large proportion of structured curation actions) tend to attract more followers on Pinterest. A similar trend is not seen
on Last.fm.
Related work
Content curation (or digital curation) is an increasingly common phenomenon. A widely accepted definition of digital curation, introduced by the UK Digital Curation Centre, reads as: “Digital curation, broadly interpreted, is about
Online social networks, especially social aggregation websites (e.g. digg, Slashdot, delicio.us and reddit), have supported the process of categorising and sharing content for
a few years. Although quantitative studies on structured curation (e.g. tags (Li, Guo, and Zhao 2008)) or unstructured
curation (e.g. likes (Sastry 2012)) are common, a comprehensive study of both kinds of content curation has not been
carried out until now.
Several dataset-backed studies have used Twitter lists as
a curation service. For instance, (Garca-Silva et al. 2012;
Greene, O’Callaghan, and Cunningham 2012; Kim et al.
2010; Yamaguchi, Amagasa, and Kitagawa 2011) explore
users’ interest based on list names or through list aggregation. (Greene et al. 2011) proposes a method to identify
members for Twitter lists on emerging topics, so that the list
could contain the key information gatekeepers and present
a balanced perspective on the story. (Ishiguro, Kimura, and
7
http://eol.org/
Takeuchi 2012) assumes that the contents of a curated list are
manually organized to fully convey the curators intentions
and use contextual features in the curation list to understand
images. However, many of these results are specific to the
setting of Twitter lists, and cannot be directly extended. Our
data-backed study of curation actions could, based on observed characteristics of curation activities, potentially help
build similar applications either in the context of Pinterest
and Last.fm, or in more contexts as well.
Summary and discussion
This paper used a quantitative analysis of several weeks
of curation actions on two different websites, Pinterest.com
and Last.fm, combined with user surveys and interviews, to
characterise the phenomenon of content curation. First we
showed that curation adds value by highlighting a different
set of items than traditional methods such as search. Next,
we discovered that collectively, the user base of each website focused most of its curation actions on a small number
of items, resulting in an extremely skewed distribution of curation activity. This could be seen as evidence of a synchronised community focusing its attention. However, our user
studies reveal that the majority of users view curation as a
personal activity, rather than a social one. Thus, synchrony
may emerge implicitly rather than as a conscious effort of
the user base.
We then examined how people curate, and proposed a
distinction between structured curation, which highlights an
item and organises it (by pinning onto a specific board or
tagging it) and unstructured curation, which simply highlights an item by liking or loving it. Our data shows that
although users differ with some preferring unstructured, and
others structured curation actions, popular items invariably
see more structured curation activity than unstructured. Using data from Last.fm, we showed that curation tends to happen soon after first contact with an item.
Finally we asked what kinds of curation behaviours attract
followers. Our data pointed to at least three factors: consistent and regular curation actions, diversity of interests, and a
preference for structured curation (in the case of Pinterest).
References
[Beagrie 2008] Beagrie, N. 2008. Digital curation for science, digital libraries, and individuals. International Journal
of Digital Curation 1(1):3–16.
[Bhargava 2009] Bhargava, R. 2009. Manifesto for the content curator: The next big social media job of the future ?
[Carr 2012] Carr, D.
2012.
A code of conduct for
content aggregators.
The New York Times.
Available from http://www.nytimes.com/2012/03/12/
business/media/guidelines-proposed-forcontent-aggregation-online.html, last accessed 23
March 2013.
[Duh et al. 2012] Duh, K.; Hirao, T.; Kimura, A.; Ishiguro,
K.; Iwata, T.; and Yeung, C. M. A. 2012. Creating stories:
Social curation of twitter messages. In Sixth International
AAAI Conference on Weblogs and Social Media.
[Garca-Silva et al. 2012] Garca-Silva, A.; Kang, J.-H.; Lerman, K.; and Corcho, O. 2012. Characterising emergent
semantics in twitter lists. In Proceedings of the 9th international conference on The Semantic Web: research and applications, ESWC’12, 530544. Berlin, Heidelberg: SpringerVerlag.
[Greene et al. 2011] Greene, D.; Reid, F.; Sheridan, G.; and
Cunningham, P. 2011. Supporting the curation of twitter
user lists. NIPS Workshop on Computational Social Science
and the Wisdom of Crowds.
[Greene, O’Callaghan, and Cunningham 2012] Greene, D.;
O’Callaghan, D.; and Cunningham, P. 2012. Identifying topical twitter communities via user list aggregation.
arXiv:1207.0017.
[Ishiguro, Kimura, and Takeuchi 2012] Ishiguro,
K.;
Kimura, A.; and Takeuchi, K. 2012. Towards automatic image understanding and mining via social curation.
In 2012 IEEE 12th International Conference on Data
Mining (ICDM), 906 –911.
[Kim et al. 2010] Kim, D.; Jo, Y.; Moon, I.-C.; and Oh, A.
2010. Analysis of twitter lists as a potential source for discovering latent characteristics of users. In ACM CHI Workshop on Microblogging.
[Li, Guo, and Zhao 2008] Li, X.; Guo, L.; and Zhao, Y. E.
2008. Tag-based social interest discovery. In Proceedings
of the 17th international conference on World Wide Web,
WWW ’08, 675684. New York, NY, USA: ACM.
[Liu 2010] Liu, S. B. 2010. The rise of curated crisis content. In Proceedings of the Information Systems for Crisis
Response and Management Conference (ISCRAM 2010).
[Ovadia 2013] Ovadia, S. 2013. Digital content curation and
why it matters to librarians. Behavioral & Social Sciences
Librarian 32(1):58–62.
[Rotman et al. 2012] Rotman, D.; Procita, K.; Hansen, D.;
Sims Parr, C.; and Preece, J. 2012. Supporting content
curation communities: The case of the encyclopedia of life.
Journal of the American Society for Information Science and
Technology.
[Sastry 2012] Sastry, N. 2012. How to tell head from tail in
user-generated content corpora. In Proceedings of the 6th
International AAAI Conference on Weblogs and Social Media.
[Shirky 2010] Shirky, C. 2010. Talk about curation. Published as on online interview with Steve Rosenbaum. Available from http://curationnationvideo.magnify.
net/video/Clay-Shirky-6, last accessed 15 Feb 2013.
[Villi 2012] Villi, M. 2012. Social curation in audience communities: UDC (user-distributed content) in the networked
media ecosystem. Participations: The international journal
of audience and reception studies, Special section: Audience
Involvement and New Production Paradigms 9(2):616–632.
[Yamaguchi, Amagasa, and Kitagawa 2011] Yamaguchi, Y.;
Amagasa, T.; and Kitagawa, H. 2011. Tag-based user topic
discovery using twitter lists. In 2011 International Conference on Advances in Social Networks Analysis and Mining
(ASONAM), 13–20.